New EfficientNetV2 paper from Google Brain uses Progressive Resizing

Researchers from Google Brain recently published their latest work EfficientNetV2 https://arxiv.org/abs/2104.00298 and while its great that it achieves a new SOTA on ImageNet, the best part imo is that they feature in the paper how they were able to use Progressive Resizing (referencing @jeremy and fastai’s DawnBench results from 2018) to train on ImageNet up to 9x faster!

As an early fastai user I remember when Jeremy first introduced this concept and while it did take them 3 years to finally catch on, its still cool to see! :slight_smile: It’s also interesting how they built upon this concept to introduce “progressive learning” which is essentially progressive resizing + adjusting regularization depending on the image size (increase regularization as images get bigger).

9 Likes